Data Visualization Basics

A well-chosen chart compresses thousands of numbers into a single image your brain can read in seconds. A bad chart hides patterns or, worse, suggests patterns that are not there. Visualization is one of the data scientist is most important skills, both for thinking through a problem and for explaining results to a non-technical audience. The choice of chart depends on what kind of data you have and what question you are asking.

A short field guide. Use a histogram or density plot for one numeric variable to see its distribution. Use a bar chart for one categorical variable to compare counts. Use a scatter plot for two numeric variables to see if they are related. Use a line chart for a numeric variable that changes over time. Use a box plot to compare distributions across groups. Avoid pie charts for anything other than a few categories with very different sizes; humans are bad at comparing angles.

Three quick rules for honest visualization. Start the y-axis at zero for bar charts to avoid exaggerating differences. Label axes clearly with units. Avoid 3D effects, which distort the data your viewer is trying to read. The deeper rule is intent: ask yourself what question the chart is meant to answer, and design the chart to answer that one question as cleanly as possible. Anything else is decoration.

Edward Tufte, a famous designer of charts, wrote that the best visualizations show the data and nothing else. Anytime your chart contains visual elements that do not help the viewer understand the data, those elements are noise. Strong data scientists keep their charts honest, focused, and unflashy.